Prediction of adsorption energies of CmHnOp (m ≤ 2, n ≤ 6, p ≤ 2) on transition metals and alloys with machine learning methods

Metal-based catalysts are widely used in many kinds of reactions, including the hydrogenation of CO2 to alcohols. Adsorption energies of key intermediates have often been used as descriptors in high-throughput catalyst screening. However, establishing machine learning models to accurately predict adsorption energies of widely spanned species is still challenging. In the present article, we explored the predictive power of sure independence screening and sparsifying operator (SISSO), multilayer perceptron regression (MLPR), random forest regression (RFR), kernel ridge regression (KRR), support vector regression (SVR), eXtreme Gradient Boosting (XGBoost), and 20 ensemble machine learning (ML) methods for adsorption energies of 57 species involved in CO2 hydrogenation to ethanol using surface features, adsorbate features, and adsorption site features. The results show that SISSO and the five base ML methods cannot furnish models with the maximum absolute error (MAX) comparable to DFT errors. On the other hand, the MAX of most two-component and three-component ensemble ML methods is less than 0.3 eV, and the KRR+MLPR+XGBoost ensemble ML model performs the best, with the mean absolute error being 0.03 eV and MAX of only 0.17 eV. Feature importance analysis reveals that the condensed local softness is the most important feature, and there are linear relations between the condensed local softness and adsorption energies of 10 C1 species on all considered surfaces. The present work shows that ensemble ML methods outperform the base ML methods for predicting adsorption energies of widely ranged species with satisfactory accuracy and deserve further studies.